Sensing and Bio-Sensing Research 43 (2024) 100632

Contents lists available at ScienceDirect

Sensing and Bio-Sensing Research

journal homepage: www.elsevier.com/locate/sbsr

Electronic nose coupled with artificial neural network for classifying of

coffee roasting profile
Suryani Dyah Astuti a, *, Ihsan Rafie Wicaksono a, Soegianto Soelistiono a, Perwira Annissa
Dyah Permatasari b, Ahmad Khalil Yaqubi c, Yunus Susilo d, Cendra Devayana Putra e,
Ardiyansyah Syahrom f
a
Department of Physics, Faculty of Science and Technology, Airlangga University, 60115 Surabaya, Indonesia
b
Department of Mathematic, Faculty of Science and Technology, Airlangga University, 60115 Surabaya, Indonesia
c
Doctoral Degree, Faculty of Science and Technology, Airlangga University, 60115 Surabaya, Indonesia
d
Faculty of Engineering, Dr Soetomo University Surabaya, Indonesia
e
Institute of Information Management, National Cheng Kung University, Tainan, Taiwan
f
Medical Devices and Technology Centre, Universiti Teknologi Malaysia, 81310 Johor, Malaysia

A R T I C L E I N F O A B S T R A C T

Keywords: Coffee known for its diverse aromas shaped by postharvest treatments, particularly the roasting process, plays a
Electronic nose pivotal role in determining the quality of the brewed beverage. This study focuses on classifying the aroma of
TGS sensors Arabica coffee beans based on roasting temperature, employing an electronic nose equipped with a TGS gas array
Roasting temperature
sensor. The classification methodology integrates deep learning through an artificial neural network (ANN),
ANN
along with a calculation analysis utilizing the Pearson correlation coefficient. Raw Robusta coffee beans were
subjected to five distinct roasting treatments (185 ◦ C, 195 ◦ C, 205 ◦ C, 215 ◦ C, and 225 ◦ C), resulting in light
roasts, light to medium roasts, medium to dark roasts, medium to dark roasts, and dark roasts. The repeatability
test affirms the TGS sensor’s reliability, exhibiting a standard deviation (STD) below 20%. Notably, the TGS 2612
and TGS 2611 sensors, dedicated to odor detection, demonstrated excellent validity with an STD below 10%
across various roasting temperatures. Classification results from deep learning cross-validation showcase
impressive accuracy: 98.2% for Light Roasts, 98.4% for Light to Medium Roasts, 98.8% for Medium Roasts,
97.8% for Medium Roasts, and 95.9% for Dark Roasts. In conclusion, this study reveals that the E-nose, utilizing
the TGS gas sensor array with deep learning analysis, effectively detects and classifies coffee types based on
roasting time with high accuracy.

1. Introduction
The most famous and expensive thing grown on plantations is coffee,
which is consumed in more than 10 million cups every year. This cof­
fee’s flavor, background, and custom all contribute to its allure. Several
varieties of coffee, including Arabica, Robusta, or a mix of these two, are
popular on a worldwide scale. According to a report by an international
coffee group, the trade in Robusta coffee is relatively lower than Arabica
coffee [1]. When compared to Robusta, which had 47.62 million bags
exported in 2021, Arabica had a total of 82.72 million bags. At the end of
2021, the trade in Arabica coffee had grown by 4.1%, and the future
looks promising. Arabica is promising both in terms of commerce overall
and price. Compared to Robusta, Arabica coffee is generally more costly.
The price differential per kg ranges from 1.26 to 2.35 USD [2].
Robusta is cultivated at lower elevations and is renowned for its
strong character, while Arabica is valued for its delicate taste and subtle
texture. Because Arabica is of more excellent quality and commands
higher prices, it dominates worldwide trade. Market impact comes from
Brazil, a significant manufacturer [3]. The sector adapts to changing
customer needs by incorporating ethical and sustainable practices.
Distinctive features, consumer inclinations, and ethical and economic
considerations all impact the dynamics of the world coffee market.
A crucial stage in making coffee is roasting, which involves giving
green coffee beans heat. The coffee is given a nice aroma by roasting,
which also removes the water from the beans. Because roasting changes
the chemistry, Arabica coffee has a more complex chemistry, as well as a

* Corresponding author.
E-mail address: suryanidyah@fst.unair.ac.id (S.D. Astuti).

https://doi.org/10.1016/j.sbsr.2024.100632
Received 28 November 2023; Received in revised form 6 February 2024; Accepted 16 February 2024
Available online 17 February 2024
2214-1804/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
S.D. Astuti et al.
better smell, taste, and color. This makes the roasting process crucial
[4]. This procedure, done at different temperatures, produces different
roast levels, giving the coffee its particular smells and scents. The beans’
chemical transformations produce a range of tastes, including Maillard
browning and caramelization, from light to dark roasts. Dark roasts
provide more profound, smokier overtones, while light roasts maintain
the natural bean qualities. To reconcile maintaining the natural char­
acteristics of the bean with reaching the appropriate roast degree,
roasting calls for dexterity and skill. Roasters of coffee manipulate
temperature and time precisely to create a harmoniously balanced taste
profile. The quality of coffee is greatly influenced by roasting. For the
coffee to taste harmonic and pleasurable, careful control over the pro­
cess is necessary to prevent under- or over-roasting. Coffee roasting is a
painstaking art form that combines scientific knowledge with sensory
perception to transform raw beans into a fragrant, tasty beverage ready
to be brewed. Roasting temperature and time determine roasting levels,
requiring a roaster to be close by to be examined the complexity of this
process accounts for the coffee roaster’s expensive compensation. The
caliber of specialists is determined by their business. According to an
official survey conducted by a third party the average salary for a coffee
roaster is between 15,000 and 20.000 USD annually, with the highest
salary surpassing 100.000 USD [5]. We are trying to come up with an
alternative to forecasting coffee roasting levels based on these expense
and complexity issues.
The Electronic Nose (E-Nose) is a revolutionary solution to the
problems with traditional coffee roasting level assessment techniques.
Conventional methods are expensive and complicated since they depend
on expert roasters and subjective human judgments. With deep learning
and gas sensors, the E-Nose provides a reliable and effective substitute. It
delivers sensitive, real-time analysis by imitating the nostrils of mam­
mals, addressing problems such as age-dependent sensory fluctuations
and subjective “cup tests.” The research shows how well the E-Nose
categorizes Arabica coffee beans according to roasting temperature,
providing a consistent and trustworthy method for assessing coffee
quality.
Therefore, to address things objectively, technology such as the
Electronic Nose (E-Nose) has been developed. E-nose has been widely
used for quality detection of food ingredients [6] and microbes that
cause infection [7] and contaminants [8]. E-nose is a device that mimics
the workings of the mammalian nostrils by using gas sensors that can
respond to certain scents [9]. The signal response generated by the e-
nose to certain scents will be analyzed using pattern recognition soft­
ware so that it can be analyzed and identified [10]. When compared
with other analytical techniques, such as gas chromatography and
electronic nose systems can be built in and can provide sensitive and
selective analysis in real time. The e-nose system has four main com­
ponents, namely the gas sensor array, headspace system, data acquisi­
tion and pattern recognition [11] [12]. The gas sensors used by E-Noses
are made of conductive polymer gas sensors, quartz microbalances,
surface acoustic vibrations, and metal oxides [13]. The sensing and
purging mechanisms make up the two main components of the head­
space system. This invention shows that it is possible to effectively
convert the chemical reaction between volatile substances and gas
sensors into a digital signal [14]. Various studies have used comparisons
of various methods to see discrimination during coffee washing [15], as
well as coffee recognition using E-nose [16] and the use of E-nose and E-
tongue to differentiate coffee in China based on species [17].
The roasting degree can be changed to suit customer preferences
[18]. Five roasting degrees were used, this research roasting level is not
indicative of the real roasting level [19]. Since level five requirements
are the most important in this study, clients with lower-level needs can
still use our method.
Coffee is classified using multispectral and random forest methods
[20]. Linear Discriminant Analysis (LDA) is used to ascertain the type of
coffee’s flavor [14]. However, all characteristics and predictors must be
assumed to have normal distributions in order to use LDA, using a
2
Sensing and Bio-Sensing Research 43 (2024) 100632
Support Vector Machine (SVM) and Random Forest (RF) [21]. However,
it is very challenging to separate data with small vector dimensions due
to the spread of data that can be obtained at a particular location.
Backpropagation analysis is used to forecast Robusta and Arabica coffee
grinds [22]. Unfortunately, extensive study is required to determine
backpropagation’s ideal performance. Principal Component Analysis
(PCA), is a technique related to Gas Chromatography-Mass Spectrom­
etry, which is used to analyze the flavor of Java and Sumatra Robusta
coffee [23]. PCA can be used with small data points, but requires a deep
learning framework to take full advantage of data.
Six gas sensors (TGS 2620, TGS 2612, TGS 2611, TGS 2602, TGS
2600, and TGS 826) are used to examine the scents produced by Arabica
coffee beans at different roasting temperatures. Every sensor provides
distinct information. Organic solvents are detected by TGS 2620, com­
bustion gases are targeted by TGS 2612, methane is detected by TGS
2611, and volatile organic chemicals are identified by TGS 2602,
cooking. Tobacco scents are targeted by TGS 2600, and ammonia and
amines are detected by TGS 826. Combined, these sensors provide a
thorough grasp of coffee odors’ subtleties and chemical makeup. This is
the foundation for a deep learning model that evaluates coffee quality
using fragrance profiles.
The model has difficulty dealing with dark roast coffee because of its
natural qualities and consistent and robust fragrance profile. Difficulty
in distinguishing fragrance qualities emerges from complicated chemi­
cal reactions caused by the lengthy, high-temperature roasting process.
Sensitivity issues with strong dark roast scents might cause the E-Nose
system’s sensors to miss more minor differences. Subjectivity in human
judgments and limited variety in the training dataset adds to the chal­
lenge. Accurate categorization is further complicated by the possible
overlap in fragrance qualities between dark roast and neighboring roast
degrees. A sophisticated strategy that includes feature inclusion, sensor
improvement, and dataset refining is needed to overcome these obsta­
cles for better dark roast categorization.
The study uses an electronic nose (E-Nose) with TGS gas sensors to
forecast coffee roasting levels, specifically Arabica coffee. The E-Nose
mimics mammalian nostrils and uses pattern recognition software for
analysis. The technology offers objectivity and is suitable for quality
detection in food ingredients [24]. The E-Nose classifies Arabica coffee
beans based on roasting temperature, determining final product char­
acteristics. The deep learning classification achieves high accuracy
values, making it a promising alternative to traditional methods for
assessing coffee quality based on aroma profiles.
The E-Nose sensor is employed to classify Arabica coffee beans based
on roasting temperature and to identify the most informative sensor for
distinguishing coffee profiles. Raw Robusta coffee beans were divided
into five samples, with each sample undergoing a different roasting
treatment: roasting at 185 ◦ C, 195 ◦ C, 205 ◦ C, 215 ◦ C, and 225 ◦ C. This
process aimed to obtain coffee bean samples representing light roasts,
light to medium roasts, medium to dark roasts, and dark roasts.
2. Materials and methods
2.1. Sample preparation
The sample is 2 kg of Robusta coffee beans from the Ijen area, East
Java, Indonesia which are still green in color and contain a moisture
content of 12.9%. Robusta coffee beans that were still raw were divided
into five samples with each sample going through a different roasting
process [25], namely roasting at 185 ◦ C, 195 ◦ C, 205 ◦ C, 215 ◦ C, and
225 ◦ C, in order to obtain samples of light roasted coffee beans (S1),
light to medium roasted coffee beans (S2), medium to dark roasted
coffee beans (S3), medium to dark roasted coffee beans (S4), and dark
roasted coffee beans (S5).
Lightly roasted Robusta coffee is produced by a mild roasting method
at 185 degrees Celsius. The moderate-to-medium roasting process is
carried out at 195 ◦ C. Roasting takes place at 205 ◦ C for the medium
S.D. Astuti et al.
roast variety. Roasting takes place at a temperature of 215 ◦ C for the
medium to dark roast. Moreover, it is roasted to 225 ◦ C during the dark
roasting stage. Data was gathered in the form of light roast, light-to-
medium roast, medium roast, medium-to-dark roast, and dark roast
coffee beans for each treatment sample, which analyzed with the E-Nose
up to 50 times.
2.2. Set-up of experiment
The experimental setting for this study is shown in Fig. 1 with an
array of gas sensors that collect Robusta coffee aromas during the data
collection phase. Previously, pre-heating or heating was carried out on
the sensor for 30 min, aiming to stabilize all sensors so that the sensor
can run well when it detects samples. Next, the sensor response was
tested with 50 replications. The output data obtained by the gas sensor
array is the value of voltage over time. Meanwhile, the input value is in
the form of the aroma obtained from the sample. During the pre-
processing phase, an array of sensors is used to determine the impor­
tance of each sensor. Computational analysis was used to predict coffee
groups using Pearson’s correlation coefficients. (See Fig. 2.)
Six sensors TGS 2620, TGS 2612, TGS 2611, TGS 2602, TGS 2600,
and TGS 826 are used in this investigation. Each sensor has a particular
gas detection range [20]. More specifically, the TGS 2620 detects vapors
of organic solvents and alcohol gases between 50 and 5000 ppm EtOH.
Methane, propane, and butane each have a detection range of 1.25% LEL
for the TGS 2612. The TGS 2611 can detect methane at levels between
500 and 10,000 ppm. TGS 2602 has a 1 to 30 ppm EtOH detection range
for volatile organic substances (hydrogen sulfide, VOC, ammonia, LPG,
butane and propane). The TGS 2600 can detect tobacco and cooking
odor gas at concentrations between 1 and 30 ppm. The TGS 826 can
detect ammonia and amine compound at concentrations between 30 and
300 ppm. Samples were observed using these sensors.
2.3. Pearson correlation analysis
Pearson correlation analysis is often used to choose which features to
use or to figure out how to change new features [26,27]. Pearson cor­
relation analysis helps to avoid overfitting and reduce computer work.
The Pearson correlation coefficient is usually used to measure the linear
relationship between two variables. Its values range from − 1 to 1. A
score of − 1 indicates a perfectly negative correlation. A number of 1
indicates a fully positive correlation, while a value of 0 indicates a linear
relationship between two variables. The Pearson correlation coefficient
is described as follows:
∑
n
(xi − x)(yi − y)
i=1
R = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (1)
∑ n ∑ n
(xi − x)2 (yi − y)2
i=1 i=1
where xi and yi stand for the X and Y values of the it sample, x and y are
the mean values of X and Y, respectively, and n is the number of samples
[24].
Fig. 1. Experiment setup.
3
Sensing and Bio-Sensing Research 43 (2024) 100632
2.4. Deep learning framework
Gathering information from E-Nose sensors when roasting Robusta
coffee beans is part of integrating a deep learning framework after
Pearson correlation analysis. Critical sensors are identified by correla­
tion analysis, and a double-hidden-layer backpropagation neural
network is built to forecast roasting levels. After training and parameter
adjustment, the model outperforms conventional techniques regarding
accuracy. The method advances the evaluation of coffee quality by of­
fering an efficient and objective way to predict coffee roasting levels
based on fragrance patterns.
Neural networks can be used to model and predict nonlinear systems,
mimicking any continuous nonlinear map. In this paper, a double-
hidden-layer backpropagation neural network with several inputs and
six outputs is built. The input system prediction parameters are thought
to be representations of the neurons found in the input layer. To com­
plete the spatially weighted aggregate of the excitation input and output
signals, the first hidden layer’s neuron nodes are employed. Neuron
nodes in the second hidden layer are used to enhance the network’s
nonlinear mapping capabilities for intricate input-output interactions.
Two process neuron nodes make up the output layer, which completes
the system output.
The transfer function of each layer has a big effect on how well a
backpropagation neural network works. Through experimentation,
several transfer functions were discovered. The linear transfer function,
tangent transfer function, logarithmic sigmoid transfer function, and
others are frequently used transfer functions in backpropagation neural
networks. Tensor Flow library defines three types of transfer functions:
tangents, linear functions, and tangents56. This can be phrased as fol­
lows, under the presumption that the signal from the input layer to the
first hidden layer is:
∑
m
mj = xk + wkj + bj (2)
k=1
where xk is the input neuron and xk is each design parameter of the Odor
Sensor discussed in this work. wkj represents the input layer to the
weight of the first hidden layer, whereas bj indicates the bias of the first
hidden layer. The output signal for the first hidden layer is indicated by
yj in Eq. (3):
( )
yj = tansig mj (3)
The signal transferred from the first hidden layer to the second
hidden layer is symbolized by the symbol ni , and its calculation is
depicted in Eq. (4):
∑
k
ni = yj wji + bi (4)
j=1
where wji is the weight of the second hidden layer’s first hidden layer
and bi is the bias of the second hidden layer. The output signal of the
second hidden layer is symbolized by the symbol zi and is expressed as
S.D. Astuti et al.
Fig. 2. Deep learning framework.
follows:
zi = tansig (ni ) (5)
The signals transferred from the second hidden layer to the output
layer are recorded as s, respectively, as illustrated in Eq. (6):
p
∑
zi = zi wi + bi (6)
i=1
2.5. Fine-tuning deep learning procedure
Depending on the amount of data needed, deep learning is often
trained using a variety of approaches. Various deep learning techniques
often result in different results. The following details our deep learning
weighting strategy: Initializing a model in step 1 is rendered absolutely
unnecessary. A model needs to be adjusted for a certain objective in
order to solve the issue. The starting node of a model typically includes a
random node weight and bias. These weights typically vary from 0 to 1.
Depending on how challenging the problem is, deep learning uses a
different amount of weight.
Forward propagation is the process of performing calculations on
data to identify the current model’s anticipated label. Step three is to
determine the loss function. To determine the current model’s perfor­
mance, a loss function computation is necessary. Generally, the mean
square error (data ratio), category cross entropy (ordinal), and other
comparable metrics serve as the foundation for the loss function. We use
categorical cross-entropy as a loss function because our dataset is
ordinal. We have classes and data, where the projected label and the
actual label of the data are both present. The following is a definition of
the total loss function:
∑
N calculations outlined for the training dataset to get the Pearson corre­
L(̂y i , yi ) = − yi loĝy i (7)
i
Fourth step is Optimization. Optimization is a technique that lowers
the overall inaccuracy of the previous step. There are several optimi­
zation techniques, including Adam, Adagrad, RMS Prop, etc.;
The Fifth step is back propagation. The backpropagation algorithm
reduces the error value in the loss function using the delta rule or
gradient descent to obtain the ideal weight. Based on the loss function,
this phase decides whether weight needs to be added or subtracted.
There are weight modifications from the last layer to the input.
Increasing both the error rate and layer weight if the derivative value is
positive. The weight must be decreased. If not, layer weight increases.
The derivative score of the preceding layer is then calculated.
4
Sensing and Bio-Sensing Research 43 (2024) 100632
wmn = wm−
n
1
− ϵ*α (8)
wmn is weight in node n and iteration m. ϵ is derivative score and α is
learning rate.
2.6. Data collection
For potential future analysis, our dataset is freely available for
download on GitHub [1]. Six sensors’ worth of raw data, representing
classes, total 26,002 bytes in the repository. Fig. 1 demonstrates the
normal distribution of the data we have gathered. The dataset has a
clustered shape rather than being divided into classes linearly. This is
consistent with our belief that natural data will result in a clustered
structure. The distribution of our dataset is balanced as well. 19,124
light roast data, 19,122 light to medium roast data, 19,132 medium
roast data, 19,508 medium to dark roast data, and 19,116 dark roast
data are all included in the dataset distribution. Any classification model
would benefit from using this condition as a data distribution. This
dataset’s intricacy is its biggest flaw. The data features cross over one
another. These conditions are difficult and require additional research to
be resolved.
3. Results
3.1. Pearson correlation coefficient
We use Pearson correlation to figure out how important each sensor
is [23,25]. This helps us find the right feature and deal with the cost
issue. We investigated whether the output of a single predictor could be
improved by integrating various olfactory descriptors. We conduct the
lation coefficient between each pair of odor descriptors. The standard
for model fusion is the Pearson correlation coefficient. The Pearson
correlation coefficients between the several odor descriptors are dis­
played in Table 1.
As the output of the regressor, each new pair of sensors is compared
to a single description. The six sensors are represented by TGS 2600 to
TGS 826. Sensor TGS 2600 and sensor TGS 2620 have a close relation­
ship. Sensor TGS 2602 and sensor TGS 826 have a close relationship.
2611 sensors and 2612 have no high relation to other sensors. Fig. 3
shows how the accuracy of predictions only changes for one or two
variables based on the output of the regressor. The most effective sensor
for predicting coffee is TGS 2602. The TGS 2600 and TGS 2620 sensors
are closely related. Despite having similar Pearson correlation
S.D. Astuti et al. Sensing and Bio-Sensing Research 43 (2024) 100632

Table 1
Pearson correlation coefficient for roasting profile coffee.
TGS 2600 TGS 2600 TGS 2602 TGS 2611 TGS 2620 TGS 2612 TGS 826

TGS 2602 0.697230 1

TGS 2611 0.411828 0.686284 1
TGS 2620 0.991888 0.735301 0.438835 1
TGS 2612 0.682535 0.498495 0.243995 0.651970 1
TGS 826 0.710215 0.972227 0.651203 0.749243 0.538378 1
Class 0.750990 0.428502 0.038299 0.776075 0.440386 0.506036

Fig. 3. Impact of descriptor correlation on prediction results.

coefficient ratings, TGS 2602 and TGS 826 have quite different per­
sonalities. This occurrence happens because, despite having identical
information, sensor TGS 2602’s information may be quieted by other
sensors, but sensor TGS 826’s information may be less quiet.
All sensors have the necessary knowledge regarding coffee classes,
just like our intuitions do. We employ all sensors as part of our features,
including the TGS 2602, TGS 2611, TGS 2612, TGS 826, TGS 2600, and
TGS 2620.
3.2. Deep learning analysis
Cross-validation was used to compare Deep Learning accuracy with
conventional techniques and represent the true state of machine
learning. The other nine groups functioned as training sets when one
group was used for testing. The ten-classification accuracy means were
applied as a consequence. Because the leave-one-out cross-validation
took longer and produced results that were identical to those of the 10-
fold cross-validation, we opted for the former in this investigation.
Cross-validation is the method we use to prevent overfitting. Node size
and hidden layer size are covered in this section. The impact of the
number of hidden layers on model performance is seen in Fig. 4. We used
the first through fourth layers. From layer 1 to layer 3, we found that
model performance improved. Layer three and the layer preceding it
experienced an increase in model generalization, which is what caused
this. Layer 4 is nonetheless less accurate than layer 3 because it is too
complex for our data. A neural network with three principal hidden
network (DHBP) layers is the last one we have.
Fig. 5 shows that the time it takes to process data goes down as the

Fig. 4. Influence hidden layer number to accuracy.

5
S.D. Astuti et al.
Fig. 5. Influence hidden layer to training time.
model gets closer to convergence. This happens because the appropriate
model only needs to adjust the weight model a few times, whereas the
inappropriate model requires more time to perform backpropagation.
This study looks at how likely it is that each layer has the same
number of nodes as shown in Fig. 6. Using 10 to 150 nodes, we assess the
model’s performance. The performance of the system is linearly influ­
enced by the number of nodes in a layer, from node 10 to node 60. The
accuracy then declines until it reaches 70. This performance degradation
is the result of a time-consuming computing process that was unable to
thoroughly evaluate the data. Performance thus improves from nodes 81
to 91. At this time, the model has learned everything it can. The model
finally reaches the maximum limit of model learning between 91 and
150 nodes, but the slow computational process makes model learning
unsatisfactory. As a result, the parameter was set to be 91.
The effect of repetition times on accuracy is depicted in Fig. 7. Each
iteration of our DHBP has comparatively higher accuracy. When the
parameters are merely set in the second iteration, the accuracy
decreases.
Other parameter settings used in this study are listed in Table 2. In
order to make this paper more reproducible, we added a parameter
table. We make use of a number of parameters, including optimizer,
epsilon, and learning rate, that are frequently utilized in deep learning
research.
Table 3 presents the classification results achieved through deep
learning on cross-validation.
Predicting odor descriptor scores using electronic nasal signal data is
an important research topic in coffee classification. Deep learning has
shown superior performance in computer vision, speech recognition,
and natural language processing compared to conventional machine
learning techniques. Table 4 shows the results of the TGS gas sensor
repeatability test. Based on the repeatability test results for each sensor,
it shows that the TGS 2612 and TGS 2611 sensors are odor sensors that
have good validity at various roasting temperatures. This discovery
Fig. 6. Influence of node number to accuracy.
6
Sensing and Bio-Sensing Research 43 (2024) 100632
could be the answer to a managerial decision to choose a cheap but
powerful odor sensor. This research found that the right number of
hidden layers affects processing time in deep learning. Technical in­
sights can be used in other domains.
Triple layers are used to overcome coffee prediction problem using E-
Nose sensors using Pearson correlation analysis and odor descriptors.
We should relax the requirement that all combinations contain the same
number of descriptors and allow combinations to choose descriptors
adaptively to get better combination effects. In the future, we will
explore other odor sensors. We also want to explore other types of
sensors to improve machine receptors such as tongue sensors.
3.3. Comparison performance across methods
According to our review, the classification of smelly coffee is closely
related to two techniques that are often used. The first approach employs
Linear Discriminant Analysis (LDA) to assess corpora [14,26]. A normal
distribution is a crucial premise for LDA. Only a large enough dataset
will allow for the normal distribution to be reached. The second tech­
nique combines Support Vector Machine (SVM) with Principal Compo­
nent Analysis (PCA) [21]. Although PCA is effective at handling
generalizations of data, it may actually result in information loss. As a
result, we used three hidden layers to improve data generalization and a
backpropagation baseline to lessen the dependence on the normal dis­
tribution. Table 3 displays our results in terms of accuracy.
In addition, we use two methods that are common in odor catego­
rization to address a variety of issues. A decision tree is the first tech­
nique [28]. When the data set is small, the decision tree or discriminant
classification method may become unstable. The first backpropagation
method is used in the second method [29]. Backpropagation is effective
in all domains, but can be improved by adding a hidden layer to capture
the essence of the smell descriptor.
Backpropagation and SVM have a 0.1 accuracy difference, but DHBP
S.D. Astuti et al. Sensing and Bio-Sensing Research 43 (2024) 100632

Fig. 7. Influence of iteration to accuracy.

Table 2
Parameters setting up of deep learning analysis.
Table 4
Notation Value Definition The result of TGS gas sensor repeatability test.
Optimizer Adam stochastic gradient descent method based on adaptive Sensor Roasting temperature
estimate of first order and second-order moments.
185 ◦ C 195 ◦ C 205 ◦ C 215 ◦ C 225 ◦ C
Epsilon 1e-07 a small positive constant that is typically used to avoid
numerical instability and divide-by-zero errors during STD (%)
gradients of a loss function calculation
TGS 2600 23.4171 8.4409621 16.1116 16.1211121 17.0965
Learning 0.001 a training parameter used to determine the weight
TGS 2602 13.8624 13.6001783 31.711 45.2457842 51.5936
Rate correction value throughout the training process
TGS 2611 5.2252 6.0196961 10.0623 8.9897177 9.658
TGS 2620 19.3861 9.0462564 17.9403 17.9323349 18.912
TGS 2612 5.1299 3.2603214 3.1297 2.7724706 3.9926
still achieves higher accuracy. The developed model is applied to fore­ TGS 826 18.3314 18.415294 36.6359 45.6665706 48.6132
casting. This model outperforms all other machine learning techniques
with a score of 0.937. We can infer from this outcome that the deep
learning approach outperformed the machine learning approach. Our
Table 5
deep learning model outperforms the competition in terms of total ac­
Comparison results of odor classifications.
curacy, coming in at 0.97 with a 0.4 accuracy difference. Table 5 shows
average number of each origin class and predicted class of cross- Author Year Method Accuracy Std.
Dev.
validation. From the table, we can notice that the most difficult profile
is dark roast. The accuracy average of deep learning is 95%. Machine Learning
[9] 2011 Support Vector Machine 0.906 0.03
Deep learning predicts light roast, medium roast, and dark roast,
[30] 2017 Random Forest 0.639 0.00
with no miss classification. The system also gets a perfect classification [18,26] 2018 and Linear Discriminant Analysis 0.706 0.00
when it comes to differentiates between light to medium roast and dark 2019
roast. The same thing also recorded in classifying the dark roast and light [27] 2020 Decision Tree 0.553 0.00
[21] 2021 Random Forest + Support Vector 0.986 0.01
to medium. Therefore, the highest accuracy for overall is achieved when
Machine
detecting the medium roast, which is 0.98. [6] 2022 Principle Component Analysis 0.987 0.01
+Deep Neural Network
4. Conclusions
Deep Learning
This research successfully used deep learning and an electronic nose [28] 2017 1-Layer Hidden Backpropagation 0.937 0.02
with TGS gas sensors to categorize Arabica coffee scents according to [31] 2022 Principle Component Analysis 0.985 0.01
+Support Vector Machine
roasting temperature. The results of the repeatability test showed that [32] 2023 Artificial neural networks 0.903 0.01
the TGS sensor used had an STD below 20% which indicated that the This 2023 Deep Learning 0.978 0.01
sensor had good validity. TGS 2612 and TGS 2611 sensors are odor study
sensors that have good validity with an STD below 10% at various

Table 3
Classification results by deep learning on cross-validation.
Odor data Predicted group membership Accuracy

Light Roast Light to Medium Roast Medium Roast Medium to Dark Roast Dark Roast

Light Roast 18,790.00 314.70 3.20 0.20 15.90 0.982

Light to Medium Roast 249.70 18,821.30 46.30 4.70 0.00 0.984
Medium Roast 2.00 20.20 18,917.30 104.50 88.00 0.988
Medium to Dark Roast 0.00 7.80 112.00 19,085.00 303.20 0.978
Dark Roast 6.20 0.00 160.10 616.00 18,333.70 0.959

7
S.D. Astuti et al.
roasting temperatures. Classification results by deep learning on cross-
validation show good accuracy values, namely 98.2% for light roasts,
98.4% for light to medium roasts, 98.8% for medium roasts, 97.8% for
Medium to dark roasts and 95.9% for dark roasts. With its objective and
effective replacement for conventional approaches, this technology has
potential uses for precise quality control and roasting level measurement
in the coffee industry. The results of this study indicate that E-nose based
on the TGS gas sensor array with deep learning analysis is able to detect
and classify coffee types based on roasting time with good accuracy.
Author statement
We the undersigned declare that this manuscript is original, has not
been published before and is not currently being considered for publi­
cation elsewhere. We confirm that the manuscript has been read and
approved by all named authors and that there are no other persons who
satisfied the criteria for authorship but are not listed. We further confirm
that the order of authors listed in the manuscript has been approved by
all of us.
CRediT authorship contribution statement
Suryani Dyah Astuti: Writing – review & editing, Visualization,
Validation, Supervision, Funding acquisition, Data curation, Conceptu­
alization. Ihsan Rafie Wicaksono: Writing – original draft, Methodol­
ogy, Investigation, Formal analysis, Data curation, Conceptualization.
Soegianto Soelistiono: Writing – original draft, Visualization, Valida­
tion, Methodology, Data curation, Conceptualization. Perwira Annissa
Dyah Permatasari: Visualization, Resources, Project administration,
Methodology, Formal analysis, Conceptualization. Ahmad Khalil
Yaqubi: Writing – review & editing, Visualization, Validation, Software,
Resources, Project administration, Formal analysis, Data curation,
Conceptualization. Yunus Susilo: Visualization, Validation, Resources,
Project administration, Methodology, Data curation, Conceptualization.
Cendra Devayana Putra: Visualization, Validation, Project adminis­
tration, Data curation, Conceptualization. Ardiyansyah Syahrom:
Visualization, Validation, Methodology, Data curation,
Conceptualization.
Declaration of competing interest
The authors state that they do not have any competing interests.
Data availability
Data will be made available on request.
References
[1] ICO, I.C.O, Exports by exporting countries to all destinations, Mon. Trade Stat. 1
(2022).
[2] M.A. Johnson, C.P. Ruiz-Diaz, N.C. Manoukis, J.C. Verle Rodrigues, Coffee berry
borer (Hypothenemus hampei), a global pest of coffee: perspectives from historical
and recent invasions, and future priorities, Insects. 11 (2010) 882.
[3] W.B. Sunarharum, D.J. Williams, H.E. Smyth, Complexity of coffee flavor: A
compositional and sensory perspective, Int. Food Res. J. 62 (2014) 315–325.
[4] N. Bhumiratana, K. Chambers Adhikari, IV, E., Evolution of sensory aroma
attributes from coffee beans to brewed coffee, LWT Food Sci. Technol. 44 (2011)
2185–2192.
[5] A. Huang, Y. Chao, E. de la Mora Velasco, A. Bilgihan, W. Wei, When artificial
intelligence meets the hospitality and tourism industry: an assessment framework
to inform theory and management, J. Hosp. Tour. 5 (2022) 1080–1100.
[6] A.I.F. Al Isyrofie, M. Kashif, A.K. Aji, N. Aidatuzzahro, A. Rahmatilah, Y. Susilo, S.
D. Astuti, Odor clustering using a gas sensor array system of chicken meat based on
temperature variations and storage time, Sens. BioSensing Res. 37 (2022) 100508.
Sensing and Bio-Sensing Research 43 (2024) 100632
[7] A.A.S. Pradhana, S.D. Astuti, M. Khasanah, R.K.D. Ardianti, Detection of gas
concentrations based on age on Staphylococcus aureus biofilms with gas array
sensors, in: AIP Conference Proceedings vol. 2314, AIP Publishing, 2020.
[8] S.D. Astuti, Y. Mukhammad, S.A.J. Duli, A.P. Putra, E.M. Setiawatie, K. Triyana,
Gas sensor array system properties for detecting bacterial biofilms, J. Med. Signals
Sens. 9 (2019) 158.
[9] E. Phaisangittisagul, H.T. Nagle, Predicting odor mixture’s responses on machine
olfaction sensors, Sensors Actuators B Chem. 155 (2011) 473–482.
[10] A.I.F. Isyrofie, A. Afifudin, R. Susilo, Y. Kholimatussa’diyah, S. Winarno, S.
D. Astuti, Role of bacterial types and odor for early detection accuracy of bacteria
with gas array, in: AIP Conf Proc 2554, 2023.
[11] S.N. Hidayat, K. Triyana, I. Fauzan, T. Julian, D. Lelono, Y. Yusuf, A.M. Peres, The
electronic nose coupled with chemometric tools for discriminating the quality of
black tea samples in situ, Chemosensors. 7 (2019) 29.
[12] J. Guo, Y. Cheng, D. Luo, K.Y. Wong, K. Hung, X. Li, ODRP: A deep learning
framework for odor descriptor rating prediction using electronic nose, IEEE Sensors
J. 21 (2021) 15012–15021.
[13] Y.H. Yu Wu, D.G. Li, H.C. Feng, Electrospun nanofibers for fast dissolution of
naproxen prepared using a coaxial process with ethanol as a shell fluid, Appl.
Mech. Mater. 662 (2014) 29–32.
[14] A. Sanaeifar, S. Mohtasebi, M. Ghasemi-Varnamkhasti, H. Ahmadi, J.S. Lozano
Rogado, Development and Application of a New Low-Cost Electronic Nose for the
Ripeness Monitoring of Banana Using Computational Techniques (PCA, LDA,
SIMCA, and SVM), 2014.
[15] S. Buratti, N. Sinelli, E. Bertone, A. Venturello, E. Casiraghi, F. Geobaldo,
Discrimination between washed Arabica, natural Arabica and Robusta coffees by
using near infrared spectroscopy, electronic nose and electronic tongue analysis,
J. Sci. Food Agric. 95 (2015) 2192–2200.
[16] K. Brudzewski, S. Osowski, A. Dwulit, Recognition of coffee using differential
electronic nose, IEEE Trans. Instrum. Meas. 61 (2012) 1803–1810.
[17] W. Dong, J. Zhao, R. Hu, Y. Dong, L. Tan, Differentiation of Chinese Robusta
coffees according to species, using a combined electronic nose and tongue, with the
aid of chemometrics, Food Chem. 229 (2017) 743–751.
[18] S.Y. Kim, B.S. Kang, A colorimetric sensor array-based classification of coffees,
Sensors Actuators B Chem. 275 (2018) 277–283.
[19] S. Romani, C. Cevoli, A. Fabbri, L. Dalla Alessandrini, M. Rosa, Evaluation of coffee
roasting degree by using electronic nose and artificial neural network for off-line
quality control, J. Food Sci. 77 (2018) (2012) C960–C965.
[20] S. Wang, K. Kirillova, X. Lehto, Travelers’ food experience sharing on social
network sites, J. Travel Tour. Mark. 34 (5) (2017) 680–693.
[21] S. Astuti, D. Tamimi, M.H. Pradhana, A.A. Alamsyah, K.A. Purnobasuki,
H. Khasanah, M. Syahrom, Gas sensor array to classify the chicken meat with E. coli
contaminant by using random forest and support vector machine, Biosens.
Bioelectron.: X 9 (2021) 100083.
[22] D. Rabersyah, Identification of types of coffee grounds using electronic nose with
backpropagation learning method, Aust. J. Electr. Electron. Eng. 5 (2016)
332–338.
[23] Y. Arimurti, K. Triyana, S. Anggrahini, Portable electronic nose as an instrument
for discrimination of Java Robusta and Sumatran Robusta coffee aromas correlated
with gas chromatography mass spectrometry, J. Phys. Sci. 10 (2018) 113–124.
[24] C.D. Putra, A.I.F. Al Isyrofie, S.D. Astuti, B.D. Putri, D.R. Ummah, M. Khasanah,
A. Syahrom, Variational autoencoder analysis gas sensor array on the preservation
process of contaminated mussel shells (Mytilus edulis), Sens. Bio-Sens. Res. 40
(2023) 100564.
[25] A.N. Gloess, A. Vietri, F. Wieland, S. Smrke, B. Schönbächler, J.A.S. López,
C. Yeretzian, Evidence of different flavour formation dynamics by roasting coffee
from different origins: on-line analysis with PTR-ToF-MS, Int. J. Mass Spectrom.
365 (2014) 324–333.
[26] F. Gottwalt, E. Chang, T. Dillon, CorrCorr: A feature selection method for
multivariate correlation network anomaly detection techniques, Comput. Secur. J
83 (2019) 234–245.
[27] G.Y.F. Makimori, E. Bona, Commercial instant coffee classification using an
electronic nose in tandem with the ComDim-LDA approach, Food Anal. Methods 12
(2019) 1067–1076.
[28] S. Wakhid, R. Sarno, S.I. Sabilla, D.B. Maghfira, Detection and classification of
Indonesian civet and non-civet coffee based on statistical analysis comparison
using E-nose, Int. J. Intell. Eng. Syst 13 (2020) 56–65.
[29] W. Zhao, Q.H. Meng, M. Zeng, P.F. Qi, Stacked sparse auto-encoders (SSAE) based
electronic nose for Chinese liquors classification, Sensors 17 (2017) 2855.
[30] A. Chemura, O. Mutanga, Developing detailed age-specific thematic maps for
coffee (Coffea Arabica L.) in heterogeneous agricultural landscapes using random
forests applied on landsat 8 multispectral sensor, Geocarto Int. 32 (2017) 759–776.
[31] S.D. Astuti, A.I.F. Al Isyrofie, R. Nashichah, M. Kashif, T. Mujiwati, Y. Susilo,
A. Syahrom, Gas array sensors based on electronic nose for detection of tuna
(Euthynnus Affinis) contaminated by Pseudomonas aeruginosa, J. Med. Signals Sens
12 (2022) 306.
[32] A.A.S. Pradhana, S.D. Astuti, P.A.D. Permatasari, R. Agustina, A.K. Yaqubi,
H. Setyawati, C.D. Putra, Sensor Array system based on electronic nose to detect
borax in meatballs with artificial neural network, J. Electr. Comput. Eng. 23 (2023)
10.